修補影片的技術在近幾年已越來越受到歡迎。大部分的技術以針對被遮蔽的物件進行處理，並通常假設背景是靜態紋理的。我們希望利用動態紋理合成的技術處理動態紋理的背景。然而，一般在動態紋理合成的方法，大多數皆為重複撥放一小段影片，並沒有考慮如何去銜接移除前後的連續性，這樣會產生一些問題。例如移除物件 後，物件 通過原本物件 通過的區塊時，如果移除整段影片，物件 便無法顯示在影片之中。又因為在影像空間中，並不容易發現動態紋理影片的規則性。這裡我們將使用流形學習來達到這個目的。首先，我們使用流形學習來描述低維度空間下的規則性，進而取代傳統線性系統的分析。這是因為動態紋理影片並不是線性的，如果能更精確的表示低維度空間的相對關係，對於重建高維影像可以得到更好的效果。此外，這裡保留物件外部周圍的影像作為參考，這樣的好處是我們可以利用保留的部份去預測消失或者是使用者想要移除的部份。一旦我們可以重建出消失或者是使用者想要移除的低維度空間座標，則可以利用這些座標來重建動態紋理。

Video inpainting or completion methods are more and more popular in recent years. Most methods focus on handling occluded objects and often assume the backgrounds are still textures. Our work desires to deal with dynamic-texture backgrounds by dynamic texture synthesis method. Most works of dynamic texture synthesis is to extent a short video into an infinite one. The connection of synthesis result is usually ignored, and it will bring some problems. For example, the user removes object , and the object pass through the path which object was appeared. Because the general way is to generate a new video to cover the object, the object will disappear in the synthesis result. In order to enhance the connection of synthesis result, we use manifolds learning to observe the correlation in low dimension space. If the coordinates in the low dimension space can be precise, the recover video can get better. Because the dynamic texture is nonlinear, the system uses nonlinear method to replace the traditional linear method. In addition, we retain a small range which outside the object. And the retain parts is used to predict the missing or the removing parts. Once the trajectory of low dimension space is predicted, the coordinates of trajectory use to recover the dynamic texture.

[1] S. Soatto, G. Doretto, and Y.-N. Wu, "Dynamic textures," in Proc. IEEE Int. Conf. Computer Vision, vol. 2, pp. 439-446, 2001.
[2] L. Yuan, F. Weng, C. Liu, and H.-Y. Shum. "Synthesizing dynamic texture with closed-loop linear dynamic system," In Proc. European Conference on Computer Vision, volume 2, pages 603-616, Prague, Czech Republic, 2004.
[3] S. Soatto, G. Doretto. "Editable Dynamic Textures," In Proc. CVPR’03, pp. 137-142, 2003.
[4] S. Roweis and L. Saul. "Nonlinear dimensionality reduction by locally linear embedding," Science, 290(5500):2323–2326, 2000.
[5] C Liu, R Lin, N Ahuja, M Yang. "Dynamic textures synthesis as nonlinear manifold learning and traversing," In The 17th British Machine Vision Conference (BMVC 2006); 859–868, 2006
[6] A. Sch¨odl, R. Szeliski, DH. Salesin, I. Essa. "Video textures," In Proceedings of SIGGRAPH 2000; 489–498, 2000
[7] A. Elgammal, C.-S. Lee. "Nonlinear manifold learning for dynamic shape and dynamic appearance," Comp. Vis. Image Underst, 2007
[8] T. Ding, M. Sznaier, and O. Camps. "A rank minimization approach to video inpainting," In Proceedings of ICCV 2007. IEEE, 2007.
[9] L.J.P. van der Maaten, E.O. Postma, and H.J. van den Herik. "Dimensionality reduction: A comparative review," IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007.
[10] L. Cayton. "Algorithms for Manifold Learning," Technical Report, 2005.
[11] Van der Maaten, L. "An introduction to dimensionality reduction using matlab," Technical Report MICC 07-07, Maastricht University, Maastricht, The Netherlands, 2007
[12] T. Poggio and F. Girosi. "Networks for approximation and learning," Proc. IEEE, vol. 78, no. 9, 1990
[13] P. Dollar, V. Rabaud, and S. Belongie, "Learning to traverse image manifolds," in Advances in Neural Information Processing Systems 19, Cambridge, MA, USA: MIT Press; pp. 361–368. ,2007
[14] Y. Bengio, J.F. Paiement, P. Vincent, and O. Delalleau. "Out-of-sample extensions for LLE, isomap, MDS, Eigenmaps, and spectral clustering," Advances in Neural Information Processing Systems, 16, 2004.
[15] J. Tenebaum, V. de Silva, and J. Langford. "A global geometric framework for nonlinear dimensionality reduction," Science, 290:2319–2323, 2000.